Loans are the core business of commercial banks, and their default risk is a critical component of banks’ risk management and asset quality monitoring. We examine the mechanisms through which CPU impacts BAQ at both the firm and bank levels. Drawing on Real Options Theory (ROT) (Abel et al., 1996), we find that CPU influences banks’ NPL by affecting firms’ investment behaviour and banks’ credit strategies. In uncertain environments, firms tend to delay investment decisions to preserve strategic flexibility (Dixit, 1989). CPU leads firms, particularly those in high-carbon industries, to postpone or scale back capital expenditures, which reduces profitability, constrains cash flows and heightens the risk of loan default. In addition, firms in high-carbon sectors face increased regulatory pressure, elevated financing costs and the risk of asset devaluation under CPU (Zhang et al., 2023), all of which amplify credit risk. This elevated risk is transmitted to banks, contributing to rising NPL. Furthermore, CPU may erode the value of firms’ collateral (e.g. high-carbon assets) and undermine managerial confidence, further impairing operational efficiency and debt-servicing capacity. In response, banks may adjust their credit structure by tightening loan approvals for high-carbon industries to mitigate potential future credit risk (Li & Wu, 2023). Nevertheless, heightened corporate credit risk may continue to increase NPL amongst banks’ existing loan portfolios.

From a systemic financial risk perspective, CPU may influence the banking sector through various risk transmission channels. Cross-shareholding and mutual investment ties between banks and traditional energy firms facilitate the transmission of stranded asset risks from the energy sector to the banking system (Zhang et al., 2023). Simultaneously, carbon emission constraints may compel the traditional energy and transportation sectors to accelerate their low-carbon transition, leading to higher operating costs and lower profits. This transition can negatively impact the value of related assets held by banks (Stroebel & Wurgler, 2021).

Based on the above discussion, we propose the following hypothesis:

The impact of CPU on BAQ may vary by bank type and size. State-owned banks, which generally possess strong capitalisation, diversified portfolios and implicit government credit guarantees (Zhang & Wang, 2020), are also highly sensitive to policy directives, including the Green Credit Guidelines, the Guiding Opinions on Building a Green Financial System and regulations pertaining to the national carbon market. As a result, under heightened CPU, this policy sensitivity may exacerbate adverse effects on their asset quality, despite structural advantages (Dong et al., 2021). In contrast, joint-stock banks, operating under a market-oriented model, enjoy greater flexibility in adjusting credit strategies and may implement more prudent lending practices to mitigate CPU-induced risks. City and rural commercial banks face greater challenges in responding to CPU due to their smaller capital buffers, less diversified credit portfolios and dependence on local industrial structures and small- and medium sized enterprises (SMEs), which increases their vulnerability to policy uncertainty (Liu et al., 2024b; Yao & Fan, 2025).

In addition, bank size plays a crucial role in determining the impact of CPU on BAQ. Larger banks benefit from stronger capital bases and diversified operations, allowing them to mitigate CPU’s adverse effects, whereas smaller banks with concentrated portfolios and limited risk-control capabilities are more susceptible to rising credit risks. (Eggers, 2020).

Based on the above discussion, we propose the following hypothesis:

H2: The impact of CPU on BAQ is different across bank type and bank size.

Banks’ CAR, corporate governance levels (CGL) and digital transformation capabilities are important factors influencing credit decisions and risk management. These elements may play a key moderating role in the relationship between CPU and BAQ.

From the perspective of banks’ risk-bearing capacity, CAR is a critical indicator of their ability to absorb losses and withstand economic shocks. A higher CAR indicates that banks have stronger capital buffers, enabling them to maintain a stable credit supply during periods of economic volatility and elevated policy uncertainty (Huang et al., 2025). Such banks typically adopt stricter credit approval and risk control mechanisms, which help reduce the generation of risky loans and mitigate the negative impact of CPU on BAQ. In addition, adequate capital reserves enhance banks’ resilience to external uncertainties, thereby decreasing the likelihood of asset quality deterioration.

From the perspective of corporate governance, the strength of a bank’s governance structure significantly affects its credit risk management capabilities. Banks with stronger governance typically have more effective internal control and shareholder oversight mechanisms, which help align managerial and shareholder interests, discourage short-termism and reduce the likelihood of excessive lending or risk-taking in response to CPU (Laeven & Levine, 2009). For example, well-governed banks typically implement institutional safeguards, including well-structured boards, transparent disclosures and rigorous risk control procedures, which promote prudent credit approval and a focus on asset quality and sustainable growth (Beltratti & Stulz, 2012). Therefore, sound corporate governance can enhance banks’ capacity to adapt to CPU, thereby mitigating its adverse impacts on BAQ.

From the perspective of information asymmetry, banks’ digital transformation capabilities directly influence their risk control abilities, which in turn determine asset quality stability under CPU. Information asymmetry is a prevalent issue in the bank credit market. During periods of high external uncertainty, the quality of corporate information disclosures may decline, hindering banks’ ability to accurately assess borrowers’ credit risk (Ng et al., 2020). However, banks with advanced digital transformation can leverage technologies such as big data analytics, artificial intelligence and machine learning to optimise credit approval processes and more accurately evaluate borrower creditworthiness (Wang et al., 2023a). This helps reduce adverse selection caused by information asymmetry and mitigates the negative impact of CPU on BAQ. Moreover, digital transformation contributes to mitigating moral hazard by allowing banks to employ intelligent monitoring systems to track borrowers’ operational performance in real time, detect emerging credit risks (Amato et al., 2024) and proactively manage default risk through tailored loan terms and liquidity provision. In a high CPU environment, banks lacking digital risk control capabilities may be more vulnerable to waves of corporate defaults triggered by policy uncertainty, further deteriorating asset quality.

Based on the above discussion, we propose the following hypothesis:

H3a: Higher CAR mitigates the negative effect of CPU on BAQ.

Based on the theoretical framework, we construct baseline regression models to test the research hypotheses presented in the previous section, computing cluster-robust standard errors at the bank level, as follows in Equation 1:

where NPL_it denotes the NPL ratio of bank i in period t, which is inversely related to BAQ, CPU_t is the CPU index, X_{mac_t} is the macroeconomic control variable, and X_{mic_it} is the microbank-level control variable. Table 1 summarises the definitions of variables.

Dependent variable: BAQ in commercial banking generally refers to a bank’s ability to recover loan principal on time and earn the expected return. It is commonly measured using indicators such as the probability of default and the ratio of NPL (Li et al., 2020; Yildirim, 2020). In this study, NPL is used as a proxy for BAQ, as a higher NPL is negatively associated with BAQ, indicating weaker asset quality.

Independent variable: We use the CPU Index developed by Ma et al. (2023) as the primary independent variable to measure climate policy changes in China (Apergis & Lau, 2015; Dai & Zhang, 2023). To align with bank-level quarterly data, we aggregate the monthly CPU into quarterly averages for regression analysis.

Control variables: To examine the impact of CPU on BAQ more comprehensively, we introduce both macrocontrol and microcontrol variables. The macroeconomic variables include the national GDP growth rate (G_GDP) and the business confidence index (BCI) to reflect real economic development. In addition, we incorporate the return on the Shanghai Composite Index (SZZS) and the Shanghai Interbank Offered Rate (SHIBOR) to account for financial market volatility, capturing capital market performance and interbank liquidity conditions. For microeconomic variables, we include several key bank-level characteristics that may influence BAQ. Return on Assets (ROA), calculated as net profit divided by total assets, serves as a measure of profitability. Bank size (Size) is represented by the natural logarithm of total assets to capture potential size-related effects. The CAR reflects the bank’s risk-bearing capacity and ability to absorb losses, whilst the CGL is proxied by equity concentration, specifically the shareholding percentage of the largest shareholder, to address potential agency problems.

Other variables: We use the Digital Transformation Index of Chinese Commercial Banks (Fintech), developed by Peking University, to measure the level of digital transformation. This index is constructed based on banks’ annual reports, patent data and other public sources, systematically capturing banks’ performance in areas such as technological infrastructure, digital service capabilities, business model innovation and management transformation.

After analysing the direct impact of CPU on BAQ, we further explore its potential transmission mechanisms. Specifically, we examine three moderating variables, namely CAR, CGL and Fintech. To assess whether CPU influences BAQ through these channels, we construct interaction-term regression models whilst maintaining the same control variables as in baseline regression. The model specification (Equation 2) is as follows:

where X_{mod_it} is the CPU interaction term for each of the three different bank metrics, that is, CAR, CGL and Fintech.

Our research sample comprises quarterly data from 29 A-share listed commercial banks in China, spanning 2008–2022. Financial information is primarily sourced from the China Securities Market and Accounting Research Database (CSMAR) and the Wind Database. To ensure the validity and reliability of our empirical results, we employ unbalanced panel data and exclude observations with excessive missing key variables. The final sample consists of 1044 quarterly observations.

Table 2 presents the descriptive statistics for all variables. NPL ranges from a minimum of 0.0038 to a maximum of 0.0312, indicating considerable variation in loan quality amongst China’s listed commercial banks. The standard deviation of ROA is 0.218, reflecting significant fluctuations in banks’ profitability. CGL has a standard deviation of 17.89, suggesting that some banks have a dominant majority shareholder, whilst others exhibit more dispersed ownership. CPU reached a low of 1.128 in Q2 2009 and a high of 3.381 in Q4 2021, with a quarterly mean of 2.231 and a standard deviation of 0.486. Figure 1 illustrates a significant upward trend in CPU over the sample period, indicating a gradual increase in CPU and volatility. This suggests that the impact of CPU on commercial banks may become increasingly complex and challenging. To examine the relationships amongst the variables, we compute the correlation coefficients, with results shown in Online Appendix 1: Table 1-A1. The analysis confirms no multicollinearity issues, as all variance inflation factors (VIF) values are below 10.

Table 3 presents the regression results of the benchmark model, which examines the impact of CPU on BAQ. According to the regression results, column (1) indicates that CPU has a statistically significant impact on NPL at the 1% level. Specifically, a one-unit increase in CPU raises NPL by 0.191 percentage points. This finding suggests that higher CPU increases credit risk and deteriorates asset quality, supporting Hypothesis H1, that is, increased CPU can hinder BAQ. Economically, the baseline effect of CPU on BAQ is non-trivial. A one-standard-deviation increase in CPU is associated with an increase in the NPL ratio of approximately 0.093 percentage points, based on the estimated coefficient of 0.191. Relative to the sample mean NPL of 1.3%, this corresponds to an increase of about 7.2%, suggesting that policy uncertainty is associated with economically meaningful changes in banks’ non-performing loans. A potential mechanism is that heightened climate policy changes lead to increased operational uncertainty for firms, arising from regulatory adjustments, carbon emission controls and rising environmental compliance costs. These factors may reduce firm profitability and elevate default risk, thereby weakening banks’ credit quality. In addition, CPU may influence banks’ risk pricing, credit approval and post-loan management, further contributing to the deterioration of asset quality.

Column (2) of Table 3 shows that the coefficients for G_GDP, SZZS and BCI are statistically significant at the 1% level, indicating that the macroeconomic environment has a notable impact on BAQ. Specifically, both G_GDP and BCI are significantly negatively correlated with NPL, suggesting that economic growth and increased business confidence contribute to improved asset quality. Gross domestic product growth is typically associated with higher corporate profitability and rising household income, which reduces loan default rates and enhances banks’ credit stability (Ghosh, 2015). Moreover, higher BCI reflects more active investment and consumption by firms and households, improved business conditions (Nowzohour & Stracca, 2020) and lower credit risk, all of which support lower NPL. Conversely, the significant positive correlation between SZZS and NPL may indicate the adverse effects of stock market volatility on BAQ. During stock market upswings, firms and individuals may increase leverage to invest in riskier financial assets (Acharya & Viswanathan, 2011), thereby heightening uncertainty regarding future debt repayment. In addition, if banks extend substantial credit to capital market-related firms, stock market fluctuations may further impair loan quality.

Column (3) of Table 3 shows that CAR is significantly and negatively correlated with NPL at the 1% level. This suggests that banks with higher CAR possess a stronger risk-absorbing capacity, enabling them to mitigate credit risks effectively, thereby reducing NPL and improving overall asset quality. Our findings align with Dai and Zhang (2023), who report that banks with higher CAR demonstrate greater robustness in risk management and proactively limit risk exposure. Similarly, ROA is significantly and negatively correlated with NPL at the 1% level, indicating that higher profitability is associated with lower credit risk. ROA measures a bank’s efficiency in generating income from its assets, with higher ROA reflecting greater profitability and more effective control of lending risk.

To test Hypothesis H2, this section conducts two heterogeneity analyses. Firstly, we examine the effect of bank type on the regression results by dividing the sample banks into state-owned banks, large joint-stock banks and urban and rural commercial banks, and conducting separate subsample regressions. Secondly, we introduce an interaction term between bank size and CPU to explore how banks of different sizes respond differently to CPU.

Table 4 presents the results of the heterogeneity analysis based on bank type and bank size. Columns (1)–(3) show that the effect of CPU on the asset quality of state-owned banks, city banks and rural commercial banks is significantly negative at the 1% level, whilst the effect on joint-stock banks is negative at the 10% level. These findings indicate that CPU has a heterogeneous impact on BAQ across different bank types. Firstly, though state-owned banks typically have strong capital bases and well-developed risk management systems, they are highly sensitive to policy directives, aligning their credit decisions closely with government priorities. Under heightened CPU, this connection may amplify the adverse effects on asset quality. Secondly, compared with national joint-stock banks, local banks such as city and rural commercial banks are more reliant on local government support and directly influenced by regional economic conditions, making their asset quality more sensitive to changes in CPU (Yao & Fan, 2025). In addition, local banks usually operate with smaller capital buffers and less sophisticated risk diversification tools, weakening their capacity to absorb shocks from policy uncertainty and external risks.

In Column (4) of Table 4, the interaction between bank size and CPU is negative and statistically significant at the 10% level, suggesting that bank size moderates the impact of CPU on BAQ. Economically, a one-standard-deviation increase in bank size (Size expressed as log (total assets)/100) increases the marginal effect of a one-standard-deviation increase in CPU on the NPL by approximately 0.00083 percentage points. Relative to the sample mean NPL of 1.3%, this implies an increase of about 6.4% relative to the sample mean, suggesting that larger banks are relatively more sensitive to policy uncertainty in terms of non-performing loans. Though the finding is consistent with the hypothesis that larger banks’ stronger capitalisation, diversified operations and advanced risk-management capabilities buffer against policy uncertainty, it does not provide conclusive evidence regarding the precise causal mechanism. A possible explanation is that larger banks respond to heightened policy uncertainty by proactively adjusting lending standards and portfolio allocations, which dampens the observable impact of CPU on NPL. Furthermore, greater regulatory scrutiny and timely supervisory interventions may partially mitigate the transmission of policy uncertainty to credit performance. Collectively, the findings offer suggestive, rather than definitive, evidence that bank size contributes to resilience against CPU-induced credit risk.

In summary, both bank type and size significantly moderate the impact of CPU on BAQ, thereby supporting Hypothesis H2.

To test Hypotheses H3, we sequentially introduce the interaction terms between CPU and three representative variables, that is, CAR, CGL and Fintech, into Model (2). Table 5 presents the regression results of Model (2) and shows that improvements in CAR, corporate governance and digital transformation significantly mitigate the adverse impact of CPU on BAQ. This indicates that banks with stronger risk absorption, more efficient governance and greater technological adaptability cope more effectively with policy uncertainty. Our findings support Hypotheses H3a, H3b and H3c.

Specifically, column (1) of Table 5 shows that the coefficient of the interaction term between CPU and CAR is –7.120 and significant at the 1% level. This indicates that a higher CAR relatively weakens the positive effect of CPU on the NPL, suggesting that stronger risk buffers relatively mitigate the impact of CPU on BAQ. Thus, Hypothesis H3a is supported. Column (2) shows that the interaction between CPU and CGL has a coefficient of –0.006, significant at the 5% level. This suggests that banks with more concentrated ownership and efficient governance relatively mitigate the negative effect of CPU on the NPL, supporting Hypothesis H3b. Column (3) reports that Fintech is significantly positively correlated with NPL at the 1% level, suggesting that increased digitalisation heightens the responsiveness of observed NPL without implying a deterioration in the underlying quality of loan portfolios. A plausible explanation is that greater Fintech adoption improves the efficiency of risk identification, monitoring and recognition, resulting in earlier and more comprehensive recognition of problem loans, which may elevate reported NPL without reflecting a deterioration in underlying asset quality. Moreover, the coefficient of the interaction term between CPU and Fintech is –0.006 and significant at the 1% level, suggesting that higher levels of digitalisation attenuate the negative impact of CPU on BAQ. Theoretically, Fintech improves banks’ information-processing capacity, facilitates real-time risk monitoring and supports adaptive credit allocation, allowing banks to respond more efficiently to shocks induced by CPU. Accordingly, increased digitalisation attenuates the negative effect of elevated CPU on BAQ, providing empirical support for Hypothesis H3c.

Economically, the heterogeneity results indicate that bank-specific characteristics substantially shape how policy uncertainty translates into credit risk. A one-standard-deviation increase in CAR (s.d. = 0.021) reduces the marginal effect of a one-standard-deviation increase in CPU on the NPL by approximately 0.073 percentage points, corresponding to about 5.6% of the sample mean NPL (1.3%). This suggests that stronger capital buffers enhance banks’ loss-absorbing capacity and dampen the transmission of policy uncertainty into asset quality deterioration. Similarly, a one-standard-deviation increase in CGL (SD = 17.890) lowers the marginal impact of CPU on NPL by around 0.052 percentage points, or roughly 4.0% of the sample mean, indicating that effective governance structures mitigate agency frictions and improve internal risk control when uncertainty rises. The moderating role of digital transformation is even more pronounced. A one-standard-deviation increase in the Fintech (s.d. = 39.950) reduces the marginal effect of CPU on NPL by approximately 0.117 percentage points, equivalent to about 9.0% of the average NPL. This pattern implies that digitalisation enhances banks’ information-processing, monitoring and adaptive response capabilities, allowing them to better withstand policy-induced uncertainty shocks.

Taken together, these results highlight that capital strength, governance quality and digital capabilities act as complementary buffers that attenuate the adverse effects of policy uncertainty on BAQ, providing strong economic support for Hypotheses H3a – H3c.

To enhance the clarity and conciseness of the main text, all robustness checks, including alternative sample period and lag structure tests, endogeneity tests and exogenous policy shock tests, are presented in Online Appendix 1. Readers seeking detailed robustness analyses are referred to the appendix.